Indeed. But the precise retiming, ie. sub-frame sampling is what could be done with the generative nodes mentioned above, no? It would be utterly memory-heavy but TimeBlur could sample the upstream with a different time delta for each pixel if need be.